[xiangshengyun]

Summary

Fix multi-GPU inference Workers receiving an unsupported model-facing device value such as cuda:0.

Problem

Each Worker is isolated to one physical GPU through CUDA_VISIBLE_DEVICES. Inside that Worker, the assigned GPU is already exposed as the local device
cuda:0. Passing DEVICE=cuda:0 to some model runtimes is not supported and can cause inference to fail with:

unsupported device request: cuda:0

## Changes

- Normalize Worker model-facing device requests:
    - cuda:N → cuda
    - cuda → cuda
    - auto → auto
    - reject unsupported Worker device values such as cpu

- Keep physical GPU isolation through CUDA_VISIBLE_DEVICES.
- Apply the normalized device value to DEVICE and SURE_EVAL_DEVICE_ACTUAL.
- Update smoke execution and parallel Worker environment handling.
- Add regression coverage for Worker device normalization.

## Validation

- Worker smoke inference completed successfully.
- Multi-GPU inference completed all 24,624 samples without the cuda:0 device error.
- Updated Worker device values are compatible with the model runtime.